KR102228552B1 - Method of predicting multiple targets of drug utilizing random forest QSAR models - Google Patents

Abstract

A method of predicting the binding of a drug to a plurality of targets present in the human body is disclosed. After generating an independent random forest prediction model for each target by machine learning the structure information of a drug that binds to the targets, the structure information of the drug to be predicted is input to the prediction model. In the random forest prediction model for each target, the possibility of binding of the predicted drug, in which the structural information for each of the targets is input, is generated as a target score for each target, and the target score for each target is calculated with the drug based on the score of the evaluation data. It converts to a drug-target binding probability, which is the probability that the target will interact. Based on the drug-target binding probability, targets having a predetermined binding probability value or more are selected and returned to the predicted target list of the predicted drug. By implementing such an algorithm in the drug target prediction server, the user can search for the highest target of the drug to be predicted through the client terminal.

Description

{Method of predicting multiple targets of drug utilizing random forest QSAR models}

The present invention relates to the field of technology for predicting a target of a drug, and more particularly, a random forest quantitative structure-activity relationship model having parameters optimized for target identification. It relates to a method for predicting multiple targets of drugs using

The human body is made up of various biological substances (eg, protein, RNA, DNA, etc.), and through the interaction of these substances, it performs metabolic activities and sustains life. When the human body suffers from a disease, abnormal activity or inactivity of certain substance(s) is observed within the human body. It is the cause or result of the disease, and can cause various symptoms of the disease. Therefore, people have developed various drugs that target specific substances in the human body, and through this, they can treat diseases or relieve symptoms.

Uncertainty about the toxicity, low efficacy and clinical safety of new drugs is a major cause of clinical failure, increasing the cost and time required to develop approved new drugs. In recent years, considerable technological progress has been made in the field of phenotypic screening technology. These recent advances provide new insights into the chemical reactions of biological networks or systems. In other words, with the development of phenotypic screening technology, it has become possible to observe the response of living systems such as cells to drugs. As a result, drugs that control the entire system, not specific targets, have become a new spotlight. However, since such drugs still exert their efficacy by binding to specific substances in the human body, it remains a major task to understand the mechanism of action by revealing the targets of drugs.

In addition, many drugs targeting a specific substance frequently fail to commercialize due to unexpected side effects. The biggest cause of the failure is the off-target effect that occurs when the drug binds to a target other than the target target. Even if the target of the drug is already known, it is necessary to predict its association with other targets. It is generally recognized that most drugs act on more than one target. If the various targets of the newly developed drug can be predicted in advance, it is possible to predict side effects in advance and optimize the drug in the direction of reducing them.

Predicting an unknown target of a drug as such is very useful in terms of pharmaceutical and biological aspects. Traditional methods of identifying drug targets include affinity chromatography, 2D gel electrophoresis, and several methods based on mRNA expression. Using these methods, drug targets can be identified with high accuracy. However, it takes a huge amount of time and money to experimentally determine which target of a new drug binds to a number of human proteins. This may cause the development cost to increase year by year in the development of new drugs.

Due to these limitations, in-silico target prediction (in a virtual environment such as computer simulation) is considered a promising alternative for target identification. In-silico target prediction can be classified into two categories: ligand-based methods and structure-based methods. In particular, the ligand-based method is advantageous in large-scale virtual screening because of its low computational cost and high feasibility. One of the most common methods of ligand-based target identification is to classify ligands using structure-activity relationships (SAR). The principle of the SAR approach is that structurally similar ligands can have similar properties. The purpose of the SAR approach is to predict the activity of the query ligand by exploring the chemical space containing the ligand structure with known activity. In in-silico target prediction, the structure of ligands can be expressed as molecular descriptors or molecular fingerprints, and activity can be defined as binding to a specific target.

Various chemogenomics libraries such as PubChem, ChEMBL, WOMBAT, and ZINC provide assay data that measures the strength of binding between various drugs and targets. Attempts to predict new targets of drugs using computational and statistical techniques based on such drug-target binding information are in the spotlight. Machine learning techniques are mainly used to effectively discover features from a large number of drug information. Representative examples of machine learning techniques are based on support-vector machine (SVM), naive Bayesian classifier (NB), artificial neural network (ANN), kernel discrimination, etc. There is a machine learning method using

Of these methods, NB is effective in classifying the target of the ligand, but is weak when there is a conditional dependence on the molecular characteristics. In addition, machine learning methods other than NB do not accurately predict the binding of a true target among a large set of more than 1,000 human targets, and report a recall rate indicating the ratio of accurately predicted targets among a specific number of predicted targets. Did not do it. Due to the nature of the technology that performs the function of target selection prior to drug experimentation, the ability to select as many correct answer targets as possible among a small number of predicted targets is important and is a measure of the efficiency of the algorithm.

Some of the issues associated with the use of SARs for target prediction are the imbalance of activity data and the ambiguity of inactive ligands across the target. This problem is based on the dependence of a ligand-based approach on the available data. Major proteins, which have been actively tested for decades, have data that are more active than other targets. In addition, in many related studies, a ligand known to be inactive against a target is considered an inactive ligand against that target. However, some of the actual ligand-target interactions may not have been tested. These biases observed in databases can lead to failures in predicting actual interactions, especially for targets with low activity data. It is believed that one of the machine learning techniques, the random forest (RF) algorithm, avoids overfitting and can properly handle unbalanced data.

In consideration of these points, the present invention is to provide a method for predicting the binding of drugs to 1,121 human targets by constructing a multi-target model using a random forest machine learning algorithm as a standardized sampling method. In particular, the present invention is based on a cross-validation result, a standard for defining inactive ligands and a comprehensive drug multi-target prediction including multiple target models by optimizing the ratio between active and inactive ligands. To build a model.

The present invention also provides a method for easily predicting the binding result of a drug to a specific target group designated by a user by generating an independent random forest prediction model for each target.

In addition, another object of the present invention is that since each target model calculates an objective probability that a drug and a target will bind, it is possible to predict the binding of a drug to a target that can present a drug-target binding probability to a user rather than a simple target ranking. Is to provide a way.

The problem to be solved by the present invention is not limited to the above-described problems, and may be variously expanded without departing from the spirit and scope of the present invention.

A method for predicting multiple targets of a drug using a random forest prediction model according to embodiments for realizing an object of the present invention is a method of predicting a drug against a plurality of targets related to the human body using a computer program executed in a computer device. A method of predicting binding, the method comprising: generating an independent random forest prediction model for each target by machine learning structural information of a drug that binds to the targets; Inputting structural information of the predicted drug into each random forest prediction model; Generating, in a random forest prediction model for each target, a binding probability of the predicted drug, into which structural information for each of the targets, is input, as a target score for each target; Converting, by the random forest prediction model for each target, the target score for each target into a drug-target binding probability, which is a probability of interaction between the drug and the target; And selecting targets having a predetermined binding probability value or more based on the drug-target binding probability and returning the predicted target list of the predicted drug.

In example embodiments, the'generating a random forest prediction model' increases the accuracy of each random forest prediction model by optimizing the data to be trained before the machine learning, It may include performing pre-processing to reduce bias between targets.

In exemplary embodiments, the performing of the pretreatment comprises a negative-undersampling method of randomly selecting only a subset of the inactive ligand until the ratio of the active ligand to the inactive ligand reaches a specific value. ) Can be used to resolve the excessive imbalance of the ratio of active ligand to inactive ligand.

In exemplary embodiments, the step of performing the pretreatment is positive-selecting relatively more active ligands by assigning a greater weight to the active ligand until the ratio of the active ligand to the inactive ligand reaches a specific value. Oversampling methods may be used to resolve excessive imbalances in the ratio of active ligand to inactive ligand.

In example embodiments, the performing of the pre-processing comprises calculating a Tanimoto coefficient Tc similarity between extended-connectivity fingerprints (ECFPs) of active and inactive ligands, Ligands with similar activity data may include performing a treatment to exclude from inactive ligands.

In example embodiments, the'generating as a target score' may include performing target filtering to exclude a specific target from among the targets from the result.

In exemplary embodiments, the method for predicting multiple targets of the drug further includes evaluating a recall rate of the method for predicting multiple targets using an external test set to verify the performance of the method for predicting multiple targets. can do.

In example embodiments, the random forest prediction model for each target may express the possibility of binding between each of 1,121 targets and the drug to be predicted as score data having a length of 1,121.

In example embodiments, the target score for each target may be converted into the drug-target binding probability based on a score of evaluation data obtained from an external validation set.

In exemplary embodiments, the inputting comprises providing a user interface through a web server computer, allowing the user to input structure information of a query ligand of a drug to be predicted through the user interface displayed on a client terminal. It may include.

In example embodiments, the method for predicting multiple targets of the drug may further include the step of providing, by the web server computer, the information on the expected target list in a web page to the client terminal.

In exemplary embodiments, the method of predicting multiple targets of a drug comprises collecting structural information of a drug that binds to each of specific targets from bioactivity data on drug-target interactions from various sources. It may contain more.

In example embodiments, the random forest prediction model may be constructed using drug activity data of a ChEMBL database for 1,121 targets.

According to exemplary embodiments of the present invention, RF QSAR, a new algorithm, may predict the binding of drugs to 1,121 human targets using a random forest machine learning technique.

When evaluated with an independent external validation set, the method for predicting drug binding to human targets according to the present invention exhibits a reproducibility that exceeds the performance of conventional NB models. In addition, due to the characteristics of the algorithm that generates an independent random forest model for each target, it is easy to predict the result for a specific target group designated by the user. In addition, since each target model calculates an objective probability that a drug and a target will bind, it is possible to present a drug-target binding probability to the user rather than a simple target ranking.

In particular, in comparison with the existing algorithm, even though the external test set, which has a lower similarity than the conventional one-which is relatively more difficult to predict-was used, a higher reproducibility was achieved and excellent performance was demonstrated. That is, for the top 11 targets (corresponding to the top 1% of all targets) and the top 33 targets (corresponding to the top 3% of all targets), it shows a recall of 66.6% and a recall of 73.9%, respectively.

1 shows an overall procedure of a method for predicting multiple targets of a drug using an RF-QSAR model executed in a server computer according to an exemplary embodiment of the present invention.
2 shows the class distribution of 1121 target models.
3 is a graph showing the relationship between the model score and the estimated probability of interaction, the left graph is a graph of the estimated probability according to the score, and the right graph is a graph of the estimated probability according to the log scale score.
4 shows the ROC curve and the region sub-curve calculated by the internal cross-validation.
5 is a scatter plot of the performance for each target model and model characteristics.
6 is a graph showing a comparison of average scores between active and inactive ligands for each target.
7 shows the change in recall for various upper k values (k = 1, 4, 8, 11, 33, 66, 88, 110) measured by internal cross-validation.
8 shows the distribution of Tanimoto coefficients of active ligands closest to specific targets.
9 shows an ROC curve for selecting the results of an external validation set.
10 is a configuration of a system for a drug target prediction server in which a multi-target prediction method based on an RF-QSAR model according to an exemplary embodiment of the present invention is implemented in a web server form to provide a web service for a query of a user's client terminal Represents.
11 illustrates a result page provided by the drug target prediction server of FIG. 10 in response to a query of a user's client terminal.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the accompanying drawings. The same reference numerals are used for the same elements in the drawings, and duplicate descriptions for the same elements are omitted.

1 shows the overall procedure of a drug multi-target prediction method (hereinafter referred to as a'multi-target prediction method') using the RF-QSAR model of the present invention. This multi-target prediction method can be implemented in a web server. Details regarding this will be described later (see descriptions of FIGS. 10 and 11).

Referring to FIG. 1, in order to execute the multi-target prediction method according to an exemplary embodiment of the present invention, first, structural information of a drug that binds to each of specific targets is collected from bioactivity (drug-target interaction) data of various sources. It can be collected (step S10). According to an exemplary embodiment, 1,121 target models may be constructed as bioactivity data of the ChEMBL database. In this case, by optimizing the data to be learned, the accuracy of each random forest prediction model can be increased, and a process can be performed to reduce a bias between targets.

Using the collected structure information of the drug, it is possible to generate a random forest prediction model learned for each target by machine learning the structure information of the drug that binds to the targets. Then, the structure information of the drug to be predicted may be input into each random forest prediction model. That is, when a user inputs a query ligand into the server, a score for the target model is calculated and a score vector may be created (S20). Specifically, in each random forest prediction model, the possibility of binding between the predicted drug and a plurality of targets to which structure information has been input may be expressed as a target score. The target score can be assigned a real value in the range of 0 to 1, for example. If target scores are assigned to all 1,121 targets, the random forest prediction model can express the binding probability between each of the 1,121 targets and the predicted drug as score data having a length of 1,121, that is, a score vector.

Then the score vector can be transformed into activation probabilities. That is, each target model may convert the target score into a drug-target binding probability based on the drug binding data (S30). The drug-target binding probability refers to the probability that the drug and the target will interact.

Based on the obtained drug-target binding probability, targets having a predetermined binding probability value or more may be selected and returned to the input predicted target list of the predicted drug (S50). Finally, the targets are ranked according to the probability for the query ligand, and can be shown to the user in order according to the ranking.

According to an exemplary embodiment, a target to be searched may be filtered by class according to a user's preference. That is, the user may exclude a specific target among targets from prediction according to convenience (S40). Through such target filtering, the number of targets to be predicted can be reduced. Targets with a higher binding probability among the predicted targets through filtering may be selected and returned to a list of expected targets of the input drug. This target filtering may be performed based on the drug-target binding probability obtained in step S30.

By evaluating the recall rate of the algorithm according to the multi-target prediction method with an external test set, it is possible to compare performance with an existing algorithm. According to the comparative evaluation, it can be confirmed that the multi-target prediction method according to the present invention exhibits excellent prediction performance.

Hereinafter, each step (S10-S50) of the multi-target prediction method will be described in more detail.

With respect to step S10, the RF-QSAR method according to an exemplary embodiment of the present invention can collect structural information of a drug that binds to a specific target from bioactivity (drug-target interaction) data from various sources. have. For example, the ChEMBL (version 20) database can be used to build active and inactive training data sets for SAR modeling. Active ligands for specific targets can be defined as molecules with activity lower than 10 μM by testing using IC50, EC50, Ki and Kd. Among the human proteins deposited in ChEMBL, for example, proteins with at least 10 known binding ligands can be selected for model development. This avoids the creation of unreliable models with insufficiently low amounts of activity data.

The selected training set corresponds to 1121 targets and 235,713 unique ligands. The number of known active ligands for each target is 10-4305. In addition, target information including class, sequence and domain can be retrieved from the ChEMBL database and further utilized by the server.

Figure 2 shows the class distribution of the target model. 1121 targets can be classified under various target classes, such as enzymes, membrane receptors, and ion channels. Since most of the targets (685) are enzymes, they can be classified by subclass of enzymes such as kinases, proteases and phosphatases. Since a small number of targets belong to several classes, the total number of targets for each class is 1143 instead of 1121 (total number of targets).

According to an exemplary implementation, the RF-QSAR method may generate a random forest prediction model using a random forest algorithm for each target by machine learning drug structure information associated with targets.

Ligand data obtained from ChEMBL are normalized using the ChemAxon standardization device. "Fragment removal", "neutralization", "explicit hydrogen removal", "Clean 2D", "Mesomerize" and "Tautomerize" options can be used. The resulting SMILES can be used, for example, to generate an ECFP_4 fingerprint (an extended connectivity fingerprint of diameter 4) with a string of 2048 bits in length, for example using the RDKit python module. Then, for each target, a ligand with known activity data is used as a positive ligand, whereas a ligand without activity data can be assumed to be a negative (inactive) ligand. After the sampling and filtering process described below, target models will be trained based on Extended-Connectivity Fingerprints (ECFP) data of active and inactive ligands using a random forest (RF) algorithm implemented in the sklearn python module. I can. Individual models can be built for each target to be used for both activity prediction and target fishing.

The random forest algorithm is known to reduce bias due to overfitting and class imbalance. Since bioactivity data obtained from ChEMBL has several grades of imbalance between activity data and inactivity data, and even between targets, the random forest classification method can effectively deal with such biases. The random forest algorithm can apply bagging and subset selection techniques to overcome the instability of the decision tree model due to hierarchical characteristics. Multiple training sets can be randomly sampled to create multiple trees, and their features can be refined based on out-of-bag cases. In an exemplary embodiment of the present invention, the number of each target model may be set to 100. A score from 0 to 1 can be defined as the percentage of the tree that determines that the query ligand is active.

According to an exemplary embodiment, it is possible to optimize the data to be trained before machine learning, thereby increasing the accuracy of each random forest prediction model and performing preprocessing to reduce a bias between targets. Through this preprocessing, class imbalance and ambiguity of inactive data can be handled before training the predictive model.

Specifically, for several targets, the ratio of the active ligand to the inactive ligand is extremely large, such as 1:23,570 days. This means that the number of active ligands is significantly less than the number of inactive ligands. This imbalance can significantly reduce the accuracy of predictions. Therefore, the class imbalance can be solved by performing the pre-processing using the following two sampling methods. According to an exemplary embodiment, only a subset of the inactive ligands can be randomly selected until the ratio of the active ligand to the inactive ligand reaches a specific value using a negative-undersampling method. According to another exemplary embodiment, an active ligand may be repeatedly selected using a positive-oversampling method. Because of its practicality, the positive-over-sampling method can result in a relatively more selection of active ligands by imposing a greater weight on the active ligand during training. These methods can resolve the excessive imbalance of the ratio of active ligand to inactive ligand. In an exemplary embodiment, it is also possible to use a common ratio across the targets so that the targets do not overcommit with multiple active ligands.

Defining inactive ligands can often be controversial as inactive ligands are relatively ambiguous compared to active ligands. Some ligands for which there is no active data may actually be active, and they should be excluded from the set of inactive ligands. By calculating the Tanimoto coefficient Tc similarity between extended-connectivity fingerprints (ECFPs) of active and inactive ligands, ligands having activity data similar to a specific threshold can be excluded from the inactive ligand.

Regarding step S20, prediction performance of each random forest prediction model may be evaluated through an internal cross-validation test on the generated random forest prediction model. According to an exemplary embodiment, in order to verify the performance of the random forest model, the prediction performance of the model for training data may be evaluated using a five-fold cross-validation method. For example, 235,713 active ligands across all targets could be divided into 5 subsets, and one subset could be set separately as a test ligand set. The remaining sets of ligands may be used as training data for developing a predictive model following the pre-processing of the data. A score can be calculated on the likelihood of binding between the test ligands and the target model. Ligands with scores above the score threshold are predicted as positive labels and others can be predicted as negative.

According to an exemplary embodiment, first, the performance of each trained model for a test set can be evaluated using a receiver-operating characteristic (ROC) curve by changing a score threshold from 0 to 1. In addition, by comparing the average score of the active ligand and the average score of the inactive ligand, it can be confirmed whether the two average values are significantly different. The ratio between the average score of the active ligand and the average score of the inactive ligand can be calculated for each target and averaged, for example, five-fold. Finally, based on the score of the targets for each ligand, 1,121 targets may be arranged in order and the targets may be ranked. And the recall can be calculated. Calculation of recall can be made assuming that the highest k values (k = 4, 7, 11, 33, 66, 88 and 110) are predicted as positive from the ranked list of targets.

You can then average the ratings over five different sets of tests. By changing the sampling method, the ratio between the number of inactive ligands and the number of active ligands, and the Tanimoto coefficient Tc similarity cutoff for inactive ligands, various target prediction models can be constructed and evaluated, from which the optimal parameters can be determined.

Thus, with the full training set of ChEMBL version 20, we can create a benchmark model using an optimized preprocessing method. However, to evaluate the benchmark model, you need an independent set of validations. Therefore, additional bioactivity data can be retrieved from ChEMBL version 21 and used as an external validation set. The outer set may contain only new ligands with at least one active target selected from target models. Ligands having the same Extended-Connectivity Fingerprints (ECFP) as in the training set may also be removed from the validation set. With the resulting 13,589 external ligands, a score matrix between the validation set and 1121 target models can be obtained. Then, the ROC curve and the area under the curve (AUC) value, and the recall for the top-k targets (k = 11 and 33, which correspond to 1% and 3% of the total number of targets, respectively) can be evaluated. The evaluated recall can be compared with the results obtained through other conventional methods.

Next, it will be described in more detail with respect to step S30 of converting the predicted score between the targets and the predicted drug (ligand) into the probability of interaction between the drug and the target.

While the predicted scores obtained through hypothetical analysis as above are useful in distinguishing between inactive and active ligands, users may want to know whether the interaction with a specific score (between the target and the predicted drug) is actually active. . In ranking targets, some ligands may be less likely to interact with higher ranked targets. According to an exemplary embodiment, in order to overcome this ambiguity, a probability estimation function for converting a prediction score of a prediction model into an interaction probability may be proposed. From hypothetical analysis of the external validation set, ligand-target pairs can be distinguished by several score cutoffs from 0 to 1. For each score cutoff, interaction pairs with a score higher than that cutoff may be maintained. The probability of interaction may be estimated based on a value obtained by dividing the number of active pairs by the total number of pairs of each cutoff.

According to an exemplary embodiment for this, an external validation set can be developed using data of a new version of ChEMBL to test the performance of a benchmark model for new ligands. The drug structure of the external validation set can be entered into each target model to obtain a score for the corresponding drug-target pair. The obtained scores (total of 13589 drugs x 1121 targets) can be used to estimate the probability value. That is, by measuring the model scores of the active drug and the inactive drug from the external evaluation data, the probability that the drug is active when the score is greater than or equal to a specific value can be measured. Since the average Tc similarity value of the external test set for the nearest ligand implemented in the benchmark model is obtained as 0.55, this value may be used as the specific value.

When the estimated probability of the interaction is obtained, as illustrated in FIG. 3, the relationship between the estimated probability and the predicted score of the model can also be expressed as a graph. In FIG. 3, a graph on the left shows an estimation probability according to a prediction score of a model, and a graph on the right shows an estimation probability according to a log scaled prediction score. The curve of the graph shown can be fitted to a sigmoid function.

The algorithm of the multi-target prediction method as described above can be implemented as a program that can be executed on a computer. According to an exemplary embodiment, the program may be implemented as a target fishing model on a web-based server. Users can access the web server computer with a client terminal and use the target prediction model. This allows you to freely search for the expected target of the query ligand. Currently, using the bioactivity data of ChEMBL version 20, it is possible to construct a random forest model with optimized parameters. PHP and jQeury can be used for web programming. The ChemAxon standardizer is implemented to standardize the SMILES format exactly as it was used in training. In addition, Open Babel software is included to convert the ligand structure into a 2D picture.

Internal verification for SAR models proposed according to an exemplary embodiment of the present invention may be performed using a five-fold cross verification procedure. The performance of the internal verification is measured using the optimized sampling method and parameters. The hypothetical screening results of the five-fold cross-validation can be used first to measure the performance of each target model. Therefore, the ROC curve of each model can be calculated by taking the average of the ROC curve from the five-fold cross-validation. The area under the ROC curve (AUC) can be evaluated to assess the performance of each target model.

Figure 4 shows the ROC curve and the overall ROC curve for the 1121 target model (left figure) and a box plot of the AUC values for the targets (right figure). In the left graph, the blue dotted line represents the ROC curve for random selection with AUC = 0.5. The red curve is the ROC curve for each target, and the black line is the overall ROC curve created using all selection data across the entire target. In the right figure, the red line represents the median value of AUC. The median AUC for total ROC is 0.97. This suggests that these models can be used to distinguish active ligands from inactive ligands with good sensitivity. The box plot shows that most models (~75%) have AUC values of 0.9 or higher. Some models (~7%) have an AUC value of less than 0.7, but those models have an AUC value greater than 0.5 and a median AUC value of 0.97.

5 is a scatter plot showing the performance of each target model and model characteristics, where a on the left represents a scatter plot of AUC values, and B on the right represents a scatter plot for the average score of the activity data. Model characteristics may include the number of active ligands (class size) and the Tc similarity between active ligands (Tc in class). Each point on the graph represents the specification of each target model. The overall trend shows that models with low performance have a small class size and low intraclass Tc.

Models with low AUC values generally have a low Tc similarity between a small number of active ligands (class size) and active ligands (Tc in class) as shown in FIG. 5A. This is probably because some of the active ligands to be cross-assayed are small and do not have any other active ligands nearby for a small target class. Since the score of the target model is used to determine the true interaction between many others, the score of the active ligands should be significantly higher than the score of the inactive ligand. To confirm this trend, a five-fold cross-validation can be used to calculate the average score of the positive and negative sets for each target. 6A shows a box plot of the average score between the active and inactive ligands for each target, and b shows the ratio distribution of the average score of the active ligand to the average score of the inactive ligand. Ratio = 10 means that the average score of the activity data is 10 times greater than the average score of the inactive ligand for the target. Measure the number of targets divided by the interval of the ratio by 1, 10, 100, 1000, 10,000, 100,000, 1,000,000, and log scale the X axis of the graph. It is observed that the average score of the negative set is nearly zero for the target model (max = 0.02), while the average score of the positive set is widely distributed with a median value of 0.64 (see Fig. 6a). Targets with a low mean score in the positive set generally have a small class size and low intraclass Tc values, similar to the trend observed in the AUC distribution (see Fig. 5b). Nevertheless, the average score of the positive set of most of the target models (99%) is at least 10 times higher than the average score of the negative set (see Fig. 6b).

The virtual selection result of each query ligand may be a score vector constructed using 1121 target models. The main application of the model according to this exemplary embodiment is to rank the targets relative to the query ligand so that users can obtain a reasonable number of test targets. Therefore, the model performance of the target ranking can be verified through cross-validation.

One of the common ways to check performance is to use the recall rate for the top-ranked target. In the method according to an exemplary embodiment of the present invention, targets ranked in the highest-k (k is the number of viable targets) are recognized as active targets for the query ligand. Reproducibility is defined as TP/(TP+FN), which is the ratio of the number of detected active targets to the actual active target. Here, TP is true positive and FN is false negative. For example, the recall can be obtained by averaging over five sets of tests during a five-fold cross-validation process. Higher reproducibility means fewer active targets are missed, resulting in better sensitivity of the model.

7 shows changes in recall for various top-k thresholds measured by internal cross-validation. The recall increases with increasing the upper-k threshold (k = 1, 4, 8, 11, 33, 66, 88, 110). However, when the top-k threshold is high, many targets that are recognized as active may actually be inactive. Also, as the number of targets to be confirmed through experiments increases, the effectiveness of the model application decreases. In fact, the recall rate changes only slightly after the top-4 thresholds. In this model, the recall rates for the top-4 and top-11 (1% of all targets) targets are 0.823 and 0.871, respectively. In view of this, in general, about 10 targets out of the total targets can be proposed as candidate targets for practicality.

Defining active and inactive targets for each target is critical to successfully modeling SARs. Depending on the sampling method (negative-undersampling and positive-oversampling), two methods can be proposed to create an active and inactive set for each target model. The ligand of the target can be sampled until the number of inactive ligands reaches a fixed ratio of the number of active ligands (arbitrarily set to 20). First, the performances of different sampling methods can be compared by calculating recall and overall AUC values (Table 1) for the top 1, 4, 8, and 11 targets. Although the negative-undersampling method performs slightly better than the positive-oversampling method in terms of overall AUC, the recall is relatively lower than that obtained using the positive-undersampling method.

Sampling method Voice-under sampling Positive-oversampling Overall ROC AUC 0.975 0.956 Top 1 Reproduction 0.534 0.549 Top 4 Reproduction 0.81 0.822 Top 8 reproduction 0.849 0.855 Top 11 reproduction 0.86 0.865

Performance comparison between negative-undersampling and positive-oversampling

In addition, since the AUC value is sufficiently high in the positive-oversampling method and the reproducibility is more important for the application of target prediction, the positive-oversampling method can be selected as a general sampling method. The positive-oversampling method recognizes more active ligands as positive compared to the negative-undersampling method with p-value = 6.39E-10 in Pearson's chi-magnetic test. Multiple positive-oversampling models can be built by varying the ratio of the number of inactive ligands to the number of active ligands between 1 and 40.

Non (inactive/active) One 10 20 30 40 Overall ROC AUC 0.961 0.956 0.956 0.955 0.955 Top 1 Reproduction 0.549 0.549 0.549 0.549 0.549 Top 4 Reproduction 0.823 0.822 0.822 0.822 0.822 Top 8 reproduction 0.857 0.856 0.855 0.855 0.855 Top 11 reproduction 0.868 0.866 0.865 0.865 0.865

Performance comparison between different ratios of the number of ligands for positive-oversampling

Table 2 shows the performance comparison between the models. The results show that a balanced ratio between active and inactive ligands yields the best reproducibility at all thresholds. The overall AUC value follows the same trend. Thus, the ratio of the number of inactive ligands to the number of active ligands can be set to 1. Pearson's chi-square test has a ratio of 1 over models with ratios of 10, 20, 30 and 40 with p-values of 7.09E-3, 7.60E-4, 6.40E-5, and 1.71E-5. It shows that the model recognized more real positivity.

Many of the inactive ligands used in the target model have not been tested experimentally for that target. Some of them may also be identified as active ligands. In particular, ligands similar to known active ligands are more likely to be active. In some cases, such an inactive ligand in the model may result in an active query being evaluated as inactive.

One way to reduce this bias is to exclude to some extent inactive ligands that are similar to active ligands. The well-known Tc similarity can be used as a cutoff for this purpose. 8 shows the distribution of Tanimoto coefficients of active ligands closest to specific targets. Referring to FIG. 8, as a result of examining the Tc similarity between active ligands closest to specific targets, 95% of the pair has a Tc similarity of 0.32 or more, and 90% of the pair has a Tc similarity of 0.5 or more.

Tc cutoff 0.3 0.5 No cutoff Overall ROC AUC 0.973 0.966 0.961 Top 1 Reproduction 0.527 0.538 0.548 Top 4 Reproduction 0.815 0.823 0.823 Top 8 reproduction 0.858 0.86 0.857 Top 11 reproduction 0.87 0.871 0.868

Performance comparison between different Tc cutoffs to exclude inactive ligands

For the different Tc similarity cutoffs (0.3, 0.5 and no cutoff), reproducibility of target ranking can be examined to obtain optimal suitability for identifying targets. The results are summarized in Table 3. The results obtained by applying the Tc cutoff values showed superior performance compared to the somewhat ambiguous results obtained by applying the Tc cutoffs 0.3 and 0.5. The AUC value increases from a Tc cutoff of 0.3, but the recall is better for a Tc cutoff of 0.5.

As mentioned earlier, a Tc cutoff of 0.5 can be selected because the recall must be more discriminating for practicality. The model with a Tc cutoff of 0.5 recognized more true positives than the case without the Tc cutoff with a p value of 1.89E-6 in the chi-square test. Thus, a benchmark model can be built using the positive-oversampling method by employing optimized parameters such as'active/inactive ratio' = 1 and'Tc cutoff' = 0.5.

The inventors developed an external validation set using data from a new version of ChEMBL to test the performance of a benchmark model on new ligands. The average Tc similarity value of the set of external tests for the nearest ligand implemented in the benchmark model was 0.55. The results of hypothetical screening of the external validation set were evaluated using the ROC curve and recall. 9 shows an ROC curve for selecting the results of an external validation set. The ROC curve was plotted by defining the known activity data as a positive set, and the area under the ROC curve was 0.89. This value is lower compared to the AUC (0.97) obtained through cross-validation. This is primarily because a greater number of active interactions are graded down to a score of zero. The ROC curve has a score of 0 for about 20% of the active ligand, while the score for 93% of the inactive ligand is zero. Active ligands with a score of 0 are not explained by the model, but are included in the outer set.

Nevertheless, the results indicate that the benchmark model's performance is still high with a value of about 0.9 for external verification. In order to verify the performance of the external verification, recall for the top-k targets was also calculated. For the top 11 (1%) targets, the recall of the outer set using the benchmark model was 67.6%. For the top 33 (3%) targets, the recall is 73.9%. This result was slightly better than the performance measured by Alexise Koutsoukas et al. using the Parzen-Rosenblatt Window-based Naive Bayesian model, and the results were 66% and 73.9% for the top 1% and 3% targets.

Reproducibility obtained using the method according to an exemplary embodiment is Laplacian modified Nave Bayes (top 1% is 63.3%, top 3% is 72.1%) or Bernoulli Nave Bayes (top 1% is 62.5%, top 3). % Is better than the recall obtained using other Nave Bayes models, such as 72.5%). The WOMBAT outer set used in these tests has an average Tc value of 0.58 of the training set, while the value of the outer set used in the test of the exemplary embodiment of the present invention is 0.55, indicating that the difficulty of the problem increases. Therefore, it can be said that the performance of the method according to the exemplary embodiment of the present invention is superior to that of the conventional methods. In addition, since the current benchmark model is a simple collection of individual target models, we expect the results to be improved through further modifications.

According to an exemplary embodiment, a web server for target prediction called RF-QSAR may be provided. Users can use RF-QSAR to identify targets of multiple query ligands at once. Each ligand is evaluated by a 1121 target model and a scoring matrix between ligands and targets can be created. The score matrix is also transformed into a random matrix, where each cell represents a probability that a ligand-target interaction will be activated. The matrix can be downloaded by link so that the user can further utilize the score matrix for other inventions. For example, a score from a target model can be used as a profile of a ligand, and the toxicity of that ligand can be predicted by that profile.

According to an exemplary embodiment, the server may provide top-k targets ranked according to their likelihood of interacting with a ligand. The value of k and the target class to be searched can be determined by the user according to the purpose of target prediction. Information and cross-references including Uniprot ID, target class, sequence, domain, and similar ligands can be provided for higher-ranking targets. The proportion of each target class of the ranked targets is also displayed, allowing the user to estimate a general target class for the query ligand.

As mentioned above, the multi-target prediction method according to the exemplary embodiment described above uses a computer program, code, instruction, or a combination of one or more of them that can be executed by various computing devices. It can be implemented in a form including (hereinafter referred to as'software'). The software may configure the processing unit to operate as desired or may instruct the processing unit independently or collectively. Software and/or data may be interpreted by a processing device or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, or computer-readable storage medium. Or may be permanently or temporarily embodyed in the device. Computer programs may be distributed over networked computer systems and stored or executed in a distributed manner.

Software and data may be stored on one or more computer-readable recording media. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

A drug target prediction server can be developed by applying an algorithm reflecting the features described above. FIG. 10 is a diagram of a multi-target prediction method based on an RF-QSAR model according to an exemplary embodiment of the present invention implemented in a drug target prediction server 100 in the form of a web server, and It shows the configuration of the system 10 for providing a web service for a query. 11 illustrates a result page provided by the drug target prediction server 100 of FIG. 10 in response to a query of a user's client terminal 200. 11 illustrates a result page provided by the web service system of FIG. 10 in response to a query of a user's client terminal.

10 and 11, the drug target prediction server 100 to which the RF-QSAR model is applied may be configured in the form of a web server that provides a web service to the user's client terminal 200 through the Internet 150. . The drug target prediction server 100 may provide a user interface screen to the client terminal 200. The user can input the structure information of the drug for which the target is to be predicted, that is, query ligand information, through the user interface screen. The drug target prediction server 100 may generate a predicted upper target list result by performing an operation for a short time. The predicted target list may be generated in the form of, for example, a web page and provided to the client terminal 200. Through the web page, the user can check the list of predicted targets of the drug structure entered by the user. The user can check the binding probability of each target and the input drug in the "Probability" column, and various biological information about other targets in another column. The user can check the prediction result by excluding desired targets among 1,121 targets at their convenience. In addition, you can obtain a list of targets of different lengths according to the purpose by adjusting the number of targets to be checked.

Several new functions can also be added to the drug target prediction server 100, such as searching for preferred targets using protein sequences and highlighting common targets repeatedly found in different query ligands.

In the above, a ligand-based SAR model including 1121 individual target models trained with human bioactivity data retrieved from the ChEMBL database using a random forest algorithm has been described. The sampling method and parameters used for data preprocessing can be carefully optimized through 5-fold cross validation to maximize the reproducibility of the top-ranked targets. Activity data of all target models can be oversampled until the ratio of the number of inactive ligands to the number of active ligands is set to 1. In addition, inactive ligands similar to active ligands having a Tc cutoff higher than 0.5 can be excluded from the modeling process. Through this process, the model according to the embodiment of the present invention can overcome the imbalance between classes or targets and avoid ambiguity of inactive ligands. The resulting target model can be used not only for predicting the activity of the ligand, but also for target fishing of the query ligand, providing a ranked list of targets.

The performance of each target model is assessed using an individual ROC curve and average score, which can indicate the strength that distinguishes active and inactive ligands. The performance of the target ranking can be verified using the recall of the top-k targets. Through external verification, recall was obtained with 67.6% in the top 1% target and 73.9% in the top 3% target. This result shows that the performance obtained in the present invention is the highest. In particular, the performance is highest even though it is a relatively difficult test set with an average Tc similarity of 0.55 with the training set.

Processes can be validated using an integrated scoring scheme that is more tailored to the probabilities of using an external data set. RF-QSAR's web interface is designed to be easy to use and can provide an intuitive result page. Users can submit multiple query ligands and check the results at once. The results page may display a ranked list of targets that are likely to interact with each other. A variety of information and cross-references can be provided to each target.

One of the features of the site is filtering targets by class. Using this function, the user can search or remove a class by specifying the target class. Users can use the server for a variety of purposes, including target prediction, ligand comparison, and profile creation.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art will be able to variously modify and change the present invention without departing from the spirit and scope of the present invention described in the following claims. You will understand that you can.

The present invention can be used to develop new drugs. In particular, it can be used to effectively and accurately predict the target of a drug to be developed.

10: Random forest model-based drug multi-target prediction service system
100: multi-target prediction server
200: client terminal

Claims (15)

As a method of predicting the binding of a drug to a plurality of targets present in the human body by using a computer program executed in a computer device,
Generating an independent random forest prediction model for each target by machine learning the structure information of the drug binding to the targets;
Inputting structural information of the predicted drug into each random forest prediction model;
Generating, in a random forest prediction model for each target, a binding probability of the predicted drug, into which structural information for each of the targets, is input, as a target score for each target;
Converting, by the random forest prediction model for each target, the target score for each target into a drug-target binding probability, which is a probability of interaction between the drug and the target; And
Multiple targets of a drug using a random forest prediction model, comprising the step of selecting targets having a predetermined binding probability value or more based on the drug-target binding probability and returning the predicted target list of the predicted drug. Prediction method. The method of claim 1, wherein the'generating a random forest prediction model' increases the accuracy of each random forest prediction model by optimizing the data to be trained before the machine learning, and targets A method for predicting multiple targets of a drug, comprising performing a pretreatment to reduce a liver bias. The method of claim 2, wherein the performing of the pretreatment comprises a negative-undersampling method in which only a subset of the inactive ligands is randomly selected until the ratio of the active ligand to the inactive ligand reaches a specific value. A method for predicting multiple targets of a drug comprising the step of resolving an excessive imbalance in the ratio of active ligands to inactive ligands by using. The method of claim 2, wherein the step of performing the pretreatment comprises assigning a greater weight to the active ligand until the ratio of the active ligand to the inactive ligand reaches a specific value, thereby selecting relatively more active ligands. A method for predicting multiple targets of a drug, comprising the step of resolving an excessive imbalance in the ratio of active ligands to inactive ligands using a method. The method of claim 2, wherein the performing of the pre-processing comprises calculating a Tanimoto coefficient Tc similarity between extended-connectivity fingerprints (ECFPs) of active and inactive ligands, and thus an activity similar to a specific threshold. A method for predicting multiple targets of a drug, comprising the step of performing a treatment in which the ligands having the data are excluded from the inactive ligands. The method of claim 1, wherein the'generating as a target score' comprises performing target filtering to exclude specific targets from the results among the targets. The multi-target prediction of a drug according to claim 1, further comprising evaluating a recall rate of the multi-target prediction method using an external test set to verify the performance of the multi-target prediction method. Way. The method of claim 1, wherein the random forest prediction model for each target expresses the possibility of binding between each of 1,121 targets and the predicted drug as score data having a length of 1,121. The method of claim 1, wherein the target score for each target is converted into the drug-target binding probability based on a score of evaluation data obtained from an external validation set. Target prediction method. The method of claim 1, wherein the inputting comprises providing a user interface through a web server computer to allow a user to input structure information of a query ligand of a drug to be predicted through the user interface displayed on a client terminal. A method for predicting multiple targets of a drug, characterized in that. The method of claim 10, further comprising the step of providing, by the web server computer, information on the expected target list in a web page to the client terminal. The method of claim 1, further comprising the step of collecting structural information of a drug binding to each of specific targets from bioactivity data on drug-target interactions from various sources. Target prediction method. The method of claim 12, wherein the random forest prediction model is constructed using drug activity data of a ChEMBL database for 1,121 targets. A computer-executable program stored in a computer-readable recording medium to perform the method for predicting multiple targets of a drug according to any one of claims 1 to 13. A computer-readable recording medium in which a computer executable program for performing the method for predicting multiple targets of a drug according to any one of claims 1 to 13 is recorded.

Images